Kubernetes at Uber with Lucy Sweet

Uber’s transition from legacy in-house platforms like Peloton to Kubernetes involves migrating millions of compute cores while maintaining continuous service availability. Success hinges on a "make-before-break" deployment pattern and enforcing strict service portability by removing dependencies on specific hosts and static network ports. A six-tier criticality system, ranging from Tier 0 infrastructure to Tier 5 internal tools, guides resource allocation and incident response. To manage massive scale across over 200 clusters, engineers utilize "Grail," a distributed graph database that correlates disparate services, network groups, and infrastructure states. Current efforts focus on migrating stateful and batch workloads, addressing Kubernetes’ limitations in node eviction through new "eviction request" primitives that allow for safer, business-logic-aware pod transitions. Lucy Sweet, a staff software engineer at Uber and lead for the Kubernetes Node Lifecycle Working Group, details these strategies for maintaining operational stability at scale.